• Cuda Implementation of CNN
  • Dependency
  • Layers
  • Model Overview
  • Params
  • result

Cuda Implementation of CNN

• This project provides several layer classes and a network class with tunable parameters
• Try to solve MNIST Dataset
• MSE loss function
• mini-batch SGD
• learning rate delay

Dependency

• CUDA Toolkit 12.2
• cuBlas -- Matrix Operation
• cuDNN v8.9 -- Convolution forward and backward, activation

Layers

• Conv2d: 28x28x1 6filters 5x5 s=1 p=2 RELU
• MaxPooling2d: 28x28x6 s=2
• Conv2d: 14x14x6 16filters 5x5 s=1 p=0 RELU
• MaxPooling2d: 10x10x16 s=2
• Dense: 160 SIGMOID
• Dense: 84 SIGMOID
• Dense: 10 SIGMOID

Model Overview

Params

• Epoch 30
• BatchSize 10
• LearningRate 0.1
• DelayRate 0.1

result

Accurate: 98.2% Running time: 1036.790s Device: 3060ti